xref: /freebsd/sys/dev/cxgbe/cxgbei/cxgbei.c (revision a3266ba2697a383d2ede56803320d941866c7e76)
1 /*-
2  * Copyright (c) 2012 Chelsio Communications, Inc.
3  * All rights reserved.
4  *
5  * Chelsio T5xx iSCSI driver
6  *
7  * Written by: Sreenivasa Honnur <shonnur@chelsio.com>
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include "opt_inet.h"
35 #include "opt_inet6.h"
36 
37 #include <sys/types.h>
38 #include <sys/param.h>
39 #include <sys/kernel.h>
40 #include <sys/ktr.h>
41 #include <sys/module.h>
42 #include <sys/systm.h>
43 
44 #ifdef TCP_OFFLOAD
45 #include <sys/errno.h>
46 #include <sys/kthread.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/mbuf.h>
51 #include <sys/lock.h>
52 #include <sys/mutex.h>
53 #include <sys/condvar.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/in_pcb.h>
57 #include <netinet/toecore.h>
58 #include <netinet/tcp_var.h>
59 #include <netinet/tcp_fsm.h>
60 
61 #include <cam/scsi/scsi_all.h>
62 #include <cam/scsi/scsi_da.h>
63 #include <cam/ctl/ctl_io.h>
64 #include <cam/ctl/ctl.h>
65 #include <cam/ctl/ctl_backend.h>
66 #include <cam/ctl/ctl_error.h>
67 #include <cam/ctl/ctl_frontend.h>
68 #include <cam/ctl/ctl_debug.h>
69 #include <cam/ctl/ctl_ha.h>
70 #include <cam/ctl/ctl_ioctl.h>
71 
72 #include <dev/iscsi/icl.h>
73 #include <dev/iscsi/iscsi_proto.h>
74 #include <dev/iscsi/iscsi_ioctl.h>
75 #include <dev/iscsi/iscsi.h>
76 #include <cam/ctl/ctl_frontend_iscsi.h>
77 
78 #include <cam/cam.h>
79 #include <cam/cam_ccb.h>
80 #include <cam/cam_xpt.h>
81 #include <cam/cam_debug.h>
82 #include <cam/cam_sim.h>
83 #include <cam/cam_xpt_sim.h>
84 #include <cam/cam_xpt_periph.h>
85 #include <cam/cam_periph.h>
86 #include <cam/cam_compat.h>
87 #include <cam/scsi/scsi_message.h>
88 
89 #include "common/common.h"
90 #include "common/t4_msg.h"
91 #include "common/t4_regs.h"	/* for PCIE_MEM_ACCESS */
92 #include "tom/t4_tom.h"
93 #include "cxgbei.h"
94 
95 static int worker_thread_count;
96 static struct cxgbei_worker_thread_softc *cwt_softc;
97 static struct proc *cxgbei_proc;
98 
99 static void
100 read_pdu_limits(struct adapter *sc, uint32_t *max_tx_data_len,
101     uint32_t *max_rx_data_len, struct ppod_region *pr)
102 {
103 	uint32_t tx_len, rx_len, r, v;
104 
105 	rx_len = t4_read_reg(sc, A_TP_PMM_RX_PAGE_SIZE);
106 	tx_len = t4_read_reg(sc, A_TP_PMM_TX_PAGE_SIZE);
107 
108 	r = t4_read_reg(sc, A_TP_PARA_REG2);
109 	rx_len = min(rx_len, G_MAXRXDATA(r));
110 	tx_len = min(tx_len, G_MAXRXDATA(r));
111 
112 	r = t4_read_reg(sc, A_TP_PARA_REG7);
113 	v = min(G_PMMAXXFERLEN0(r), G_PMMAXXFERLEN1(r));
114 	rx_len = min(rx_len, v);
115 	tx_len = min(tx_len, v);
116 
117 	/*
118 	 * AHS is not supported by the kernel so we'll not account for
119 	 * it either in our PDU len -> data segment len conversions.
120 	 */
121 	rx_len -= ISCSI_BHS_SIZE + ISCSI_HEADER_DIGEST_SIZE +
122 	    ISCSI_DATA_DIGEST_SIZE;
123 	tx_len -= ISCSI_BHS_SIZE + ISCSI_HEADER_DIGEST_SIZE +
124 	    ISCSI_DATA_DIGEST_SIZE;
125 
126 	/*
127 	 * DDP can place only 4 pages for a single PDU.  A single
128 	 * request might use larger pages than the smallest page size,
129 	 * but that cannot be guaranteed.  Assume the smallest DDP
130 	 * page size for this limit.
131 	 */
132 	rx_len = min(rx_len, 4 * (1U << pr->pr_page_shift[0]));
133 
134 	if (chip_id(sc) == CHELSIO_T5) {
135 		tx_len = min(tx_len, 15360);
136 
137 		rx_len = rounddown2(rx_len, 512);
138 		tx_len = rounddown2(tx_len, 512);
139 	}
140 
141 	*max_tx_data_len = tx_len;
142 	*max_rx_data_len = rx_len;
143 }
144 
145 /*
146  * Initialize the software state of the iSCSI ULP driver.
147  *
148  * ENXIO means firmware didn't set up something that it was supposed to.
149  */
150 static int
151 cxgbei_init(struct adapter *sc, struct cxgbei_data *ci)
152 {
153 	struct sysctl_oid *oid;
154 	struct sysctl_oid_list *children;
155 	struct ppod_region *pr;
156 	uint32_t r;
157 	int rc;
158 
159 	MPASS(sc->vres.iscsi.size > 0);
160 	MPASS(ci != NULL);
161 
162 	pr = &ci->pr;
163 	r = t4_read_reg(sc, A_ULP_RX_ISCSI_PSZ);
164 	rc = t4_init_ppod_region(pr, &sc->vres.iscsi, r, "iSCSI page pods");
165 	if (rc != 0) {
166 		device_printf(sc->dev,
167 		    "%s: failed to initialize the iSCSI page pod region: %u.\n",
168 		    __func__, rc);
169 		return (rc);
170 	}
171 
172 	r = t4_read_reg(sc, A_ULP_RX_ISCSI_TAGMASK);
173 	r &= V_ISCSITAGMASK(M_ISCSITAGMASK);
174 	if (r != pr->pr_tag_mask) {
175 		/*
176 		 * Recent firmwares are supposed to set up the iSCSI tagmask
177 		 * but we'll do it ourselves it the computed value doesn't match
178 		 * what's in the register.
179 		 */
180 		device_printf(sc->dev,
181 		    "tagmask 0x%08x does not match computed mask 0x%08x.\n", r,
182 		    pr->pr_tag_mask);
183 		t4_set_reg_field(sc, A_ULP_RX_ISCSI_TAGMASK,
184 		    V_ISCSITAGMASK(M_ISCSITAGMASK), pr->pr_tag_mask);
185 	}
186 
187 	read_pdu_limits(sc, &ci->max_tx_data_len, &ci->max_rx_data_len, pr);
188 
189 	sysctl_ctx_init(&ci->ctx);
190 	oid = device_get_sysctl_tree(sc->dev);	/* dev.t5nex.X */
191 	children = SYSCTL_CHILDREN(oid);
192 
193 	oid = SYSCTL_ADD_NODE(&ci->ctx, children, OID_AUTO, "iscsi",
194 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "iSCSI ULP settings");
195 	children = SYSCTL_CHILDREN(oid);
196 
197 	ci->ddp_threshold = 2048;
198 	SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "ddp_threshold",
199 	    CTLFLAG_RW, &ci->ddp_threshold, 0, "Rx zero copy threshold");
200 
201 	SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "max_rx_data_len",
202 	    CTLFLAG_RD, &ci->max_rx_data_len, 0,
203 	    "Maximum receive data segment length");
204 	SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "max_tx_data_len",
205 	    CTLFLAG_RD, &ci->max_tx_data_len, 0,
206 	    "Maximum transmit data segment length");
207 
208 	return (0);
209 }
210 
211 static int
212 do_rx_iscsi_hdr(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
213 {
214 	struct adapter *sc = iq->adapter;
215 	struct cpl_iscsi_hdr *cpl = mtod(m, struct cpl_iscsi_hdr *);
216 	u_int tid = GET_TID(cpl);
217 	struct toepcb *toep = lookup_tid(sc, tid);
218 	struct icl_pdu *ip;
219 	struct icl_cxgbei_pdu *icp;
220 	uint16_t len_ddp = be16toh(cpl->pdu_len_ddp);
221 	uint16_t len = be16toh(cpl->len);
222 
223 	M_ASSERTPKTHDR(m);
224 	MPASS(m->m_pkthdr.len == len + sizeof(*cpl));
225 
226 	ip = icl_cxgbei_new_pdu(M_NOWAIT);
227 	if (ip == NULL)
228 		CXGBE_UNIMPLEMENTED("PDU allocation failure");
229 	m_copydata(m, sizeof(*cpl), ISCSI_BHS_SIZE, (caddr_t)ip->ip_bhs);
230 	ip->ip_data_len = G_ISCSI_PDU_LEN(len_ddp) - len;
231 	icp = ip_to_icp(ip);
232 	icp->icp_seq = ntohl(cpl->seq);
233 	icp->icp_flags = ICPF_RX_HDR;
234 
235 	/* This is the start of a new PDU.  There should be no old state. */
236 	MPASS(toep->ulpcb2 == NULL);
237 	toep->ulpcb2 = icp;
238 
239 #if 0
240 	CTR5(KTR_CXGBE, "%s: tid %u, cpl->len %u, pdu_len_ddp 0x%04x, icp %p",
241 	    __func__, tid, len, len_ddp, icp);
242 #endif
243 
244 	m_freem(m);
245 	return (0);
246 }
247 
248 static int
249 do_rx_iscsi_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
250 {
251 	struct adapter *sc = iq->adapter;
252 	struct cpl_iscsi_data *cpl =  mtod(m, struct cpl_iscsi_data *);
253 	u_int tid = GET_TID(cpl);
254 	struct toepcb *toep = lookup_tid(sc, tid);
255 	struct icl_cxgbei_pdu *icp = toep->ulpcb2;
256 	struct icl_pdu *ip;
257 
258 	M_ASSERTPKTHDR(m);
259 	MPASS(m->m_pkthdr.len == be16toh(cpl->len) + sizeof(*cpl));
260 
261 	if (icp == NULL) {
262 		/*
263 		 * T6 completion enabled, start of a new pdu. Header
264 		 * will come in completion CPL.
265 		 */
266 	        ip = icl_cxgbei_new_pdu(M_NOWAIT);
267 	        if (ip == NULL)
268 			CXGBE_UNIMPLEMENTED("PDU allocation failure");
269 		icp = ip_to_icp(ip);
270 	} else {
271 		/* T5 mode, header is already received. */
272 		MPASS(icp->icp_flags == ICPF_RX_HDR);
273 		MPASS(icp->ip.ip_data_mbuf == NULL);
274 		MPASS(icp->ip.ip_data_len == m->m_pkthdr.len - sizeof(*cpl));
275 	}
276 
277 	/* Trim the cpl header from mbuf. */
278 	m_adj(m, sizeof(*cpl));
279 
280 	icp->icp_flags |= ICPF_RX_FLBUF;
281 	icp->ip.ip_data_mbuf = m;
282 	toep->ofld_rxq->rx_iscsi_fl_pdus++;
283 	toep->ofld_rxq->rx_iscsi_fl_octets += m->m_pkthdr.len;
284 
285 	/*
286 	 * For T6, save the icp for further processing in the
287 	 * completion handler.
288 	 */
289 	if (icp->icp_flags == ICPF_RX_FLBUF) {
290 		MPASS(toep->ulpcb2 == NULL);
291 		toep->ulpcb2 = icp;
292 	}
293 
294 #if 0
295 	CTR4(KTR_CXGBE, "%s: tid %u, cpl->len %u, icp %p", __func__, tid,
296 	    be16toh(cpl->len), icp);
297 #endif
298 
299 	return (0);
300 }
301 
302 static int
303 do_rx_iscsi_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
304 {
305 	struct adapter *sc = iq->adapter;
306 	const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
307 	u_int tid = GET_TID(cpl);
308 	struct toepcb *toep = lookup_tid(sc, tid);
309 	struct inpcb *inp = toep->inp;
310 	struct socket *so;
311 	struct sockbuf *sb;
312 	struct tcpcb *tp;
313 	struct icl_cxgbei_conn *icc;
314 	struct icl_conn *ic;
315 	struct icl_cxgbei_pdu *icp = toep->ulpcb2;
316 	struct icl_pdu *ip;
317 	u_int pdu_len, val;
318 	struct epoch_tracker et;
319 
320 	MPASS(m == NULL);
321 
322 	/* Must already be assembling a PDU. */
323 	MPASS(icp != NULL);
324 	MPASS(icp->icp_flags & ICPF_RX_HDR);	/* Data is optional. */
325 	MPASS((icp->icp_flags & ICPF_RX_STATUS) == 0);
326 
327 	pdu_len = be16toh(cpl->len);	/* includes everything. */
328 	val = be32toh(cpl->ddpvld);
329 
330 #if 0
331 	CTR5(KTR_CXGBE,
332 	    "%s: tid %u, cpl->len %u, ddpvld 0x%08x, icp_flags 0x%08x",
333 	    __func__, tid, pdu_len, val, icp->icp_flags);
334 #endif
335 
336 	icp->icp_flags |= ICPF_RX_STATUS;
337 	ip = &icp->ip;
338 	if (val & F_DDP_PADDING_ERR)
339 		icp->icp_flags |= ICPF_PAD_ERR;
340 	if (val & F_DDP_HDRCRC_ERR)
341 		icp->icp_flags |= ICPF_HCRC_ERR;
342 	if (val & F_DDP_DATACRC_ERR)
343 		icp->icp_flags |= ICPF_DCRC_ERR;
344 	if (val & F_DDP_PDU && ip->ip_data_mbuf == NULL) {
345 		MPASS((icp->icp_flags & ICPF_RX_FLBUF) == 0);
346 		MPASS(ip->ip_data_len > 0);
347 		icp->icp_flags |= ICPF_RX_DDP;
348 		toep->ofld_rxq->rx_iscsi_ddp_pdus++;
349 		toep->ofld_rxq->rx_iscsi_ddp_octets += ip->ip_data_len;
350 	}
351 
352 	INP_WLOCK(inp);
353 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
354 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
355 		    __func__, tid, pdu_len, inp->inp_flags);
356 		INP_WUNLOCK(inp);
357 		icl_cxgbei_conn_pdu_free(NULL, ip);
358 		toep->ulpcb2 = NULL;
359 		return (0);
360 	}
361 
362 	/*
363 	 * T6+ does not report data PDUs received via DDP without F
364 	 * set.  This can result in gaps in the TCP sequence space.
365 	 */
366 	tp = intotcpcb(inp);
367 	MPASS(chip_id(sc) >= CHELSIO_T6 || icp->icp_seq == tp->rcv_nxt);
368 	tp->rcv_nxt = icp->icp_seq + pdu_len;
369 	tp->t_rcvtime = ticks;
370 
371 	/*
372 	 * Don't update the window size or return credits since RX
373 	 * flow control is disabled.
374 	 */
375 
376 	so = inp->inp_socket;
377 	sb = &so->so_rcv;
378 	SOCKBUF_LOCK(sb);
379 
380 	icc = toep->ulpcb;
381 	if (__predict_false(icc == NULL || sb->sb_state & SBS_CANTRCVMORE)) {
382 		CTR5(KTR_CXGBE,
383 		    "%s: tid %u, excess rx (%d bytes), icc %p, sb_state 0x%x",
384 		    __func__, tid, pdu_len, icc, sb->sb_state);
385 		SOCKBUF_UNLOCK(sb);
386 		INP_WUNLOCK(inp);
387 
388 		CURVNET_SET(so->so_vnet);
389 		NET_EPOCH_ENTER(et);
390 		INP_WLOCK(inp);
391 		tp = tcp_drop(tp, ECONNRESET);
392 		if (tp)
393 			INP_WUNLOCK(inp);
394 		NET_EPOCH_EXIT(et);
395 		CURVNET_RESTORE();
396 
397 		icl_cxgbei_conn_pdu_free(NULL, ip);
398 		toep->ulpcb2 = NULL;
399 		return (0);
400 	}
401 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
402 	ic = &icc->ic;
403 	icl_cxgbei_new_pdu_set_conn(ip, ic);
404 
405 	MPASS(m == NULL); /* was unused, we'll use it now. */
406 	m = sbcut_locked(sb, sbused(sb)); /* XXXNP: toep->sb_cc accounting? */
407 	if (__predict_false(m != NULL)) {
408 		int len = m_length(m, NULL);
409 
410 		/*
411 		 * PDUs were received before the tid transitioned to ULP mode.
412 		 * Convert them to icl_cxgbei_pdus and send them to ICL before
413 		 * the PDU in icp/ip.
414 		 */
415 		CTR3(KTR_CXGBE, "%s: tid %u, %u bytes in so_rcv", __func__, tid,
416 		    len);
417 
418 		/* XXXNP: needs to be rewritten. */
419 		if (len == sizeof(struct iscsi_bhs) || len == 4 + sizeof(struct
420 		    iscsi_bhs)) {
421 			struct icl_cxgbei_pdu *icp0;
422 			struct icl_pdu *ip0;
423 
424 			ip0 = icl_cxgbei_new_pdu(M_NOWAIT);
425 			if (ip0 == NULL)
426 				CXGBE_UNIMPLEMENTED("PDU allocation failure");
427 			icl_cxgbei_new_pdu_set_conn(ip0, ic);
428 			icp0 = ip_to_icp(ip0);
429 			icp0->icp_seq = 0; /* XXX */
430 			icp0->icp_flags = ICPF_RX_HDR | ICPF_RX_STATUS;
431 			m_copydata(m, 0, sizeof(struct iscsi_bhs), (void *)ip0->ip_bhs);
432 			STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip0, ip_next);
433 		}
434 		m_freem(m);
435 	}
436 
437 	STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip, ip_next);
438 	if ((icc->rx_flags & RXF_ACTIVE) == 0) {
439 		struct cxgbei_worker_thread_softc *cwt = &cwt_softc[icc->cwt];
440 
441 		mtx_lock(&cwt->cwt_lock);
442 		icc->rx_flags |= RXF_ACTIVE;
443 		TAILQ_INSERT_TAIL(&cwt->rx_head, icc, rx_link);
444 		if (cwt->cwt_state == CWT_SLEEPING) {
445 			cwt->cwt_state = CWT_RUNNING;
446 			cv_signal(&cwt->cwt_cv);
447 		}
448 		mtx_unlock(&cwt->cwt_lock);
449 	}
450 	SOCKBUF_UNLOCK(sb);
451 	INP_WUNLOCK(inp);
452 
453 	toep->ulpcb2 = NULL;
454 
455 	return (0);
456 }
457 
458 static int
459 do_rx_iscsi_cmp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
460 {
461 	struct epoch_tracker et;
462 	struct adapter *sc = iq->adapter;
463 	struct cpl_rx_iscsi_cmp *cpl = mtod(m, struct cpl_rx_iscsi_cmp *);
464 	u_int tid = GET_TID(cpl);
465 	struct toepcb *toep = lookup_tid(sc, tid);
466 	struct icl_cxgbei_pdu *icp = toep->ulpcb2;
467 	struct icl_pdu *ip;
468 	struct cxgbei_cmp *cmp;
469 	struct inpcb *inp = toep->inp;
470 #ifdef INVARIANTS
471 	uint16_t len = be16toh(cpl->len);
472 #endif
473 	struct socket *so;
474 	struct sockbuf *sb;
475 	struct tcpcb *tp;
476 	struct icl_cxgbei_conn *icc;
477 	struct icl_conn *ic;
478 	struct iscsi_bhs_data_out *bhsdo;
479 	u_int val = be32toh(cpl->ddpvld);
480 	u_int npdus, pdu_len, data_digest_len, hdr_digest_len;
481 	uint32_t prev_seg_len;
482 
483 	M_ASSERTPKTHDR(m);
484 	MPASS(m->m_pkthdr.len == len + sizeof(*cpl));
485 
486 	if ((val & F_DDP_PDU) == 0) {
487 		MPASS(icp != NULL);
488 		MPASS((icp->icp_flags & ICPF_RX_STATUS) == 0);
489 		ip = &icp->ip;
490 	}
491 
492 	if (icp == NULL) {
493 		/* T6 completion enabled, start of a new PDU. */
494 		ip = icl_cxgbei_new_pdu(M_NOWAIT);
495 		if (ip == NULL)
496 			CXGBE_UNIMPLEMENTED("PDU allocation failure");
497 		icp = ip_to_icp(ip);
498 	}
499 	pdu_len = G_ISCSI_PDU_LEN(be16toh(cpl->pdu_len_ddp));
500 
501 #if 0
502 	CTR5(KTR_CXGBE,
503 	    "%s: tid %u, cpl->len %u, ddpvld 0x%08x, icp %p",
504 	    __func__, tid, pdu_len, val, icp);
505 #endif
506 
507 	/* Copy header */
508 	m_copydata(m, sizeof(*cpl), ISCSI_BHS_SIZE, (caddr_t)ip->ip_bhs);
509 	bhsdo = (struct iscsi_bhs_data_out *)ip->ip_bhs;
510 	ip->ip_data_len = bhsdo->bhsdo_data_segment_len[0] << 16 |
511 	    bhsdo->bhsdo_data_segment_len[1] << 8 |
512 	    bhsdo->bhsdo_data_segment_len[2];
513 	icp->icp_seq = ntohl(cpl->seq);
514 	icp->icp_flags |= ICPF_RX_HDR;
515 	icp->icp_flags |= ICPF_RX_STATUS;
516 
517 	if (val & F_DDP_PADDING_ERR)
518 		icp->icp_flags |= ICPF_PAD_ERR;
519 	if (val & F_DDP_HDRCRC_ERR)
520 		icp->icp_flags |= ICPF_HCRC_ERR;
521 	if (val & F_DDP_DATACRC_ERR)
522 		icp->icp_flags |= ICPF_DCRC_ERR;
523 
524 	INP_WLOCK(inp);
525 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
526 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
527 		    __func__, tid, pdu_len, inp->inp_flags);
528 		INP_WUNLOCK(inp);
529 		icl_cxgbei_conn_pdu_free(NULL, ip);
530 		toep->ulpcb2 = NULL;
531 		m_freem(m);
532 		return (0);
533 	}
534 
535 	tp = intotcpcb(inp);
536 
537 	/*
538 	 * If icc is NULL, the connection is being closed in
539 	 * icl_cxgbei_conn_close(), just drop this data.
540 	 */
541 	icc = toep->ulpcb;
542 	if (__predict_false(icc == NULL)) {
543 		CTR4(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes), icc %p",
544 		    __func__, tid, pdu_len, icc);
545 
546 		/*
547 		 * Update rcv_nxt so the sequence number of the FIN
548 		 * doesn't appear wrong.
549 		 */
550 		tp->rcv_nxt = icp->icp_seq + pdu_len;
551 		tp->t_rcvtime = ticks;
552 		INP_WUNLOCK(inp);
553 
554 		icl_cxgbei_conn_pdu_free(NULL, ip);
555 		toep->ulpcb2 = NULL;
556 		m_freem(m);
557 		return (0);
558 	}
559 
560 	data_digest_len = (icc->ulp_submode & ULP_CRC_DATA) ?
561 	    ISCSI_DATA_DIGEST_SIZE : 0;
562 	hdr_digest_len = (icc->ulp_submode & ULP_CRC_HEADER) ?
563 	    ISCSI_HEADER_DIGEST_SIZE : 0;
564 	MPASS(roundup2(ip->ip_data_len, 4) == pdu_len - len - data_digest_len);
565 
566 	if (val & F_DDP_PDU && ip->ip_data_mbuf == NULL) {
567 		MPASS((icp->icp_flags & ICPF_RX_FLBUF) == 0);
568 		MPASS(ip->ip_data_len > 0);
569 		icp->icp_flags |= ICPF_RX_DDP;
570 		bhsdo = (struct iscsi_bhs_data_out *)ip->ip_bhs;
571 
572 		switch (ip->ip_bhs->bhs_opcode & ~ISCSI_BHS_OPCODE_IMMEDIATE) {
573 		case ISCSI_BHS_OPCODE_SCSI_DATA_IN:
574 			cmp = cxgbei_find_cmp(icc,
575 			    be32toh(bhsdo->bhsdo_initiator_task_tag));
576 			break;
577 		case ISCSI_BHS_OPCODE_SCSI_DATA_OUT:
578 			cmp = cxgbei_find_cmp(icc,
579 			    be32toh(bhsdo->bhsdo_target_transfer_tag));
580 			break;
581 		default:
582 			__assert_unreachable();
583 		}
584 		MPASS(cmp != NULL);
585 
586 		/*
587 		 * The difference between the end of the last burst
588 		 * and the offset of the last PDU in this burst is
589 		 * the additional data received via DDP.
590 		 */
591 		prev_seg_len = be32toh(bhsdo->bhsdo_buffer_offset) -
592 		    cmp->next_buffer_offset;
593 
594 		if (prev_seg_len != 0) {
595 			uint32_t orig_datasn;
596 
597 			/*
598 			 * Return a "large" PDU representing the burst
599 			 * of PDUs.  Adjust the offset and length of
600 			 * this PDU to represent the entire burst.
601 			 */
602 			ip->ip_data_len += prev_seg_len;
603 			bhsdo->bhsdo_data_segment_len[2] = ip->ip_data_len;
604 			bhsdo->bhsdo_data_segment_len[1] = ip->ip_data_len >> 8;
605 			bhsdo->bhsdo_data_segment_len[0] = ip->ip_data_len >> 16;
606 			bhsdo->bhsdo_buffer_offset =
607 			    htobe32(cmp->next_buffer_offset);
608 
609 			orig_datasn = htobe32(bhsdo->bhsdo_datasn);
610 			npdus = orig_datasn - cmp->last_datasn;
611 			bhsdo->bhsdo_datasn = htobe32(cmp->last_datasn + 1);
612 			cmp->last_datasn = orig_datasn;
613 			ip->ip_additional_pdus = npdus - 1;
614 		} else {
615 			MPASS(htobe32(bhsdo->bhsdo_datasn) ==
616 			    cmp->last_datasn + 1);
617 			npdus = 1;
618 			cmp->last_datasn = htobe32(bhsdo->bhsdo_datasn);
619 		}
620 
621 		cmp->next_buffer_offset += ip->ip_data_len;
622 		toep->ofld_rxq->rx_iscsi_ddp_pdus += npdus;
623 		toep->ofld_rxq->rx_iscsi_ddp_octets += ip->ip_data_len;
624 	} else {
625 		MPASS(icp->icp_flags & (ICPF_RX_FLBUF));
626 		MPASS(ip->ip_data_len == ip->ip_data_mbuf->m_pkthdr.len);
627 	}
628 
629 	tp->rcv_nxt = icp->icp_seq + pdu_len;
630 	tp->t_rcvtime = ticks;
631 
632 	/*
633 	 * Don't update the window size or return credits since RX
634 	 * flow control is disabled.
635 	 */
636 
637 	so = inp->inp_socket;
638 	sb = &so->so_rcv;
639 	SOCKBUF_LOCK(sb);
640 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
641 		CTR5(KTR_CXGBE,
642 		    "%s: tid %u, excess rx (%d bytes), icc %p, sb_state 0x%x",
643 		    __func__, tid, pdu_len, icc, sb->sb_state);
644 		SOCKBUF_UNLOCK(sb);
645 		INP_WUNLOCK(inp);
646 
647 		CURVNET_SET(so->so_vnet);
648 		NET_EPOCH_ENTER(et);
649 		INP_WLOCK(inp);
650 		tp = tcp_drop(tp, ECONNRESET);
651 		if (tp != NULL)
652 			INP_WUNLOCK(inp);
653 		NET_EPOCH_EXIT(et);
654 		CURVNET_RESTORE();
655 
656 		icl_cxgbei_conn_pdu_free(NULL, ip);
657 		toep->ulpcb2 = NULL;
658 		m_freem(m);
659 		return (0);
660 	}
661 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
662 	ic = &icc->ic;
663 	icl_cxgbei_new_pdu_set_conn(ip, ic);
664 
665 	/* Enqueue the PDU to the received pdus queue. */
666 	STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip, ip_next);
667 	if ((icc->rx_flags & RXF_ACTIVE) == 0) {
668 		struct cxgbei_worker_thread_softc *cwt = &cwt_softc[icc->cwt];
669 
670 		mtx_lock(&cwt->cwt_lock);
671 		icc->rx_flags |= RXF_ACTIVE;
672 		TAILQ_INSERT_TAIL(&cwt->rx_head, icc, rx_link);
673 		if (cwt->cwt_state == CWT_SLEEPING) {
674 			cwt->cwt_state = CWT_RUNNING;
675 			cv_signal(&cwt->cwt_cv);
676 		}
677 		mtx_unlock(&cwt->cwt_lock);
678 	}
679 	SOCKBUF_UNLOCK(sb);
680 	INP_WUNLOCK(inp);
681 
682 	toep->ulpcb2 = NULL;
683 	m_freem(m);
684 
685 	return (0);
686 }
687 
688 static int
689 cxgbei_activate(struct adapter *sc)
690 {
691 	struct cxgbei_data *ci;
692 	int rc;
693 
694 	ASSERT_SYNCHRONIZED_OP(sc);
695 
696 	if (uld_active(sc, ULD_ISCSI)) {
697 		KASSERT(0, ("%s: iSCSI offload already enabled on adapter %p",
698 		    __func__, sc));
699 		return (0);
700 	}
701 
702 	if (sc->iscsicaps == 0 || sc->vres.iscsi.size == 0) {
703 		device_printf(sc->dev,
704 		    "not iSCSI offload capable, or capability disabled.\n");
705 		return (ENOSYS);
706 	}
707 
708 	/* per-adapter softc for iSCSI */
709 	ci = malloc(sizeof(*ci), M_CXGBE, M_ZERO | M_WAITOK);
710 	if (ci == NULL)
711 		return (ENOMEM);
712 
713 	rc = cxgbei_init(sc, ci);
714 	if (rc != 0) {
715 		free(ci, M_CXGBE);
716 		return (rc);
717 	}
718 
719 	sc->iscsi_ulp_softc = ci;
720 
721 	return (0);
722 }
723 
724 static int
725 cxgbei_deactivate(struct adapter *sc)
726 {
727 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
728 
729 	ASSERT_SYNCHRONIZED_OP(sc);
730 
731 	if (ci != NULL) {
732 		sysctl_ctx_free(&ci->ctx);
733 		t4_free_ppod_region(&ci->pr);
734 		free(ci, M_CXGBE);
735 		sc->iscsi_ulp_softc = NULL;
736 	}
737 
738 	return (0);
739 }
740 
741 static void
742 cxgbei_activate_all(struct adapter *sc, void *arg __unused)
743 {
744 
745 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4isact") != 0)
746 		return;
747 
748 	/* Activate iSCSI if any port on this adapter has IFCAP_TOE enabled. */
749 	if (sc->offload_map && !uld_active(sc, ULD_ISCSI))
750 		(void) t4_activate_uld(sc, ULD_ISCSI);
751 
752 	end_synchronized_op(sc, 0);
753 }
754 
755 static void
756 cxgbei_deactivate_all(struct adapter *sc, void *arg __unused)
757 {
758 
759 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4isdea") != 0)
760 		return;
761 
762 	if (uld_active(sc, ULD_ISCSI))
763 	    (void) t4_deactivate_uld(sc, ULD_ISCSI);
764 
765 	end_synchronized_op(sc, 0);
766 }
767 
768 static struct uld_info cxgbei_uld_info = {
769 	.uld_id = ULD_ISCSI,
770 	.activate = cxgbei_activate,
771 	.deactivate = cxgbei_deactivate,
772 };
773 
774 static void
775 cwt_main(void *arg)
776 {
777 	struct cxgbei_worker_thread_softc *cwt = arg;
778 	struct icl_cxgbei_conn *icc = NULL;
779 	struct icl_conn *ic;
780 	struct icl_pdu *ip;
781 	struct sockbuf *sb;
782 	STAILQ_HEAD(, icl_pdu) rx_pdus = STAILQ_HEAD_INITIALIZER(rx_pdus);
783 
784 	MPASS(cwt != NULL);
785 
786 	mtx_lock(&cwt->cwt_lock);
787 	MPASS(cwt->cwt_state == 0);
788 	cwt->cwt_state = CWT_RUNNING;
789 	cv_signal(&cwt->cwt_cv);
790 
791 	while (__predict_true(cwt->cwt_state != CWT_STOP)) {
792 		cwt->cwt_state = CWT_RUNNING;
793 		while ((icc = TAILQ_FIRST(&cwt->rx_head)) != NULL) {
794 			TAILQ_REMOVE(&cwt->rx_head, icc, rx_link);
795 			mtx_unlock(&cwt->cwt_lock);
796 
797 			ic = &icc->ic;
798 			sb = &ic->ic_socket->so_rcv;
799 
800 			SOCKBUF_LOCK(sb);
801 			MPASS(icc->rx_flags & RXF_ACTIVE);
802 			if (__predict_true(!(sb->sb_state & SBS_CANTRCVMORE))) {
803 				MPASS(STAILQ_EMPTY(&rx_pdus));
804 				STAILQ_SWAP(&icc->rcvd_pdus, &rx_pdus, icl_pdu);
805 				SOCKBUF_UNLOCK(sb);
806 
807 				/* Hand over PDUs to ICL. */
808 				while ((ip = STAILQ_FIRST(&rx_pdus)) != NULL) {
809 					STAILQ_REMOVE_HEAD(&rx_pdus, ip_next);
810 					ic->ic_receive(ip);
811 				}
812 
813 				SOCKBUF_LOCK(sb);
814 				MPASS(STAILQ_EMPTY(&rx_pdus));
815 			}
816 			MPASS(icc->rx_flags & RXF_ACTIVE);
817 			if (STAILQ_EMPTY(&icc->rcvd_pdus) ||
818 			    __predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
819 				icc->rx_flags &= ~RXF_ACTIVE;
820 			} else {
821 				/*
822 				 * More PDUs were received while we were busy
823 				 * handing over the previous batch to ICL.
824 				 * Re-add this connection to the end of the
825 				 * queue.
826 				 */
827 				mtx_lock(&cwt->cwt_lock);
828 				TAILQ_INSERT_TAIL(&cwt->rx_head, icc,
829 				    rx_link);
830 				mtx_unlock(&cwt->cwt_lock);
831 			}
832 			SOCKBUF_UNLOCK(sb);
833 
834 			mtx_lock(&cwt->cwt_lock);
835 		}
836 
837 		/* Inner loop doesn't check for CWT_STOP, do that first. */
838 		if (__predict_false(cwt->cwt_state == CWT_STOP))
839 			break;
840 		cwt->cwt_state = CWT_SLEEPING;
841 		cv_wait(&cwt->cwt_cv, &cwt->cwt_lock);
842 	}
843 
844 	MPASS(TAILQ_FIRST(&cwt->rx_head) == NULL);
845 	mtx_assert(&cwt->cwt_lock, MA_OWNED);
846 	cwt->cwt_state = CWT_STOPPED;
847 	cv_signal(&cwt->cwt_cv);
848 	mtx_unlock(&cwt->cwt_lock);
849 	kthread_exit();
850 }
851 
852 static int
853 start_worker_threads(void)
854 {
855 	int i, rc;
856 	struct cxgbei_worker_thread_softc *cwt;
857 
858 	worker_thread_count = min(mp_ncpus, 32);
859 	cwt_softc = malloc(worker_thread_count * sizeof(*cwt), M_CXGBE,
860 	    M_WAITOK | M_ZERO);
861 
862 	MPASS(cxgbei_proc == NULL);
863 	for (i = 0, cwt = &cwt_softc[0]; i < worker_thread_count; i++, cwt++) {
864 		mtx_init(&cwt->cwt_lock, "cwt lock", NULL, MTX_DEF);
865 		cv_init(&cwt->cwt_cv, "cwt cv");
866 		TAILQ_INIT(&cwt->rx_head);
867 		rc = kproc_kthread_add(cwt_main, cwt, &cxgbei_proc, NULL, 0, 0,
868 		    "cxgbei", "%d", i);
869 		if (rc != 0) {
870 			printf("cxgbei: failed to start thread #%d/%d (%d)\n",
871 			    i + 1, worker_thread_count, rc);
872 			mtx_destroy(&cwt->cwt_lock);
873 			cv_destroy(&cwt->cwt_cv);
874 			bzero(cwt, sizeof(*cwt));
875 			if (i == 0) {
876 				free(cwt_softc, M_CXGBE);
877 				worker_thread_count = 0;
878 
879 				return (rc);
880 			}
881 
882 			/* Not fatal, carry on with fewer threads. */
883 			worker_thread_count = i;
884 			rc = 0;
885 			break;
886 		}
887 
888 		/* Wait for thread to start before moving on to the next one. */
889 		mtx_lock(&cwt->cwt_lock);
890 		while (cwt->cwt_state == 0)
891 			cv_wait(&cwt->cwt_cv, &cwt->cwt_lock);
892 		mtx_unlock(&cwt->cwt_lock);
893 	}
894 
895 	MPASS(cwt_softc != NULL);
896 	MPASS(worker_thread_count > 0);
897 	return (0);
898 }
899 
900 static void
901 stop_worker_threads(void)
902 {
903 	int i;
904 	struct cxgbei_worker_thread_softc *cwt = &cwt_softc[0];
905 
906 	MPASS(worker_thread_count >= 0);
907 
908 	for (i = 0, cwt = &cwt_softc[0]; i < worker_thread_count; i++, cwt++) {
909 		mtx_lock(&cwt->cwt_lock);
910 		MPASS(cwt->cwt_state == CWT_RUNNING ||
911 		    cwt->cwt_state == CWT_SLEEPING);
912 		cwt->cwt_state = CWT_STOP;
913 		cv_signal(&cwt->cwt_cv);
914 		do {
915 			cv_wait(&cwt->cwt_cv, &cwt->cwt_lock);
916 		} while (cwt->cwt_state != CWT_STOPPED);
917 		mtx_unlock(&cwt->cwt_lock);
918 		mtx_destroy(&cwt->cwt_lock);
919 		cv_destroy(&cwt->cwt_cv);
920 	}
921 	free(cwt_softc, M_CXGBE);
922 }
923 
924 /* Select a worker thread for a connection. */
925 u_int
926 cxgbei_select_worker_thread(struct icl_cxgbei_conn *icc)
927 {
928 	struct adapter *sc = icc->sc;
929 	struct toepcb *toep = icc->toep;
930 	u_int i, n;
931 
932 	n = worker_thread_count / sc->sge.nofldrxq;
933 	if (n > 0)
934 		i = toep->vi->pi->port_id * n + arc4random() % n;
935 	else
936 		i = arc4random() % worker_thread_count;
937 
938 	CTR3(KTR_CXGBE, "%s: tid %u, cwt %u", __func__, toep->tid, i);
939 
940 	return (i);
941 }
942 
943 static int
944 cxgbei_mod_load(void)
945 {
946 	int rc;
947 
948 	t4_register_cpl_handler(CPL_ISCSI_HDR, do_rx_iscsi_hdr);
949 	t4_register_cpl_handler(CPL_ISCSI_DATA, do_rx_iscsi_data);
950 	t4_register_cpl_handler(CPL_RX_ISCSI_DDP, do_rx_iscsi_ddp);
951 	t4_register_cpl_handler(CPL_RX_ISCSI_CMP, do_rx_iscsi_cmp);
952 
953 	rc = start_worker_threads();
954 	if (rc != 0)
955 		return (rc);
956 
957 	rc = t4_register_uld(&cxgbei_uld_info);
958 	if (rc != 0) {
959 		stop_worker_threads();
960 		return (rc);
961 	}
962 
963 	t4_iterate(cxgbei_activate_all, NULL);
964 
965 	return (rc);
966 }
967 
968 static int
969 cxgbei_mod_unload(void)
970 {
971 
972 	t4_iterate(cxgbei_deactivate_all, NULL);
973 
974 	if (t4_unregister_uld(&cxgbei_uld_info) == EBUSY)
975 		return (EBUSY);
976 
977 	stop_worker_threads();
978 
979 	t4_register_cpl_handler(CPL_ISCSI_HDR, NULL);
980 	t4_register_cpl_handler(CPL_ISCSI_DATA, NULL);
981 	t4_register_cpl_handler(CPL_RX_ISCSI_DDP, NULL);
982 	t4_register_cpl_handler(CPL_RX_ISCSI_CMP, NULL);
983 
984 	return (0);
985 }
986 #endif
987 
988 static int
989 cxgbei_modevent(module_t mod, int cmd, void *arg)
990 {
991 	int rc = 0;
992 
993 #ifdef TCP_OFFLOAD
994 	switch (cmd) {
995 	case MOD_LOAD:
996 		rc = cxgbei_mod_load();
997 		if (rc == 0)
998 			rc = icl_cxgbei_mod_load();
999 		break;
1000 
1001 	case MOD_UNLOAD:
1002 		rc = icl_cxgbei_mod_unload();
1003 		if (rc == 0)
1004 			rc = cxgbei_mod_unload();
1005 		break;
1006 
1007 	default:
1008 		rc = EINVAL;
1009 	}
1010 #else
1011 	printf("cxgbei: compiled without TCP_OFFLOAD support.\n");
1012 	rc = EOPNOTSUPP;
1013 #endif
1014 
1015 	return (rc);
1016 }
1017 
1018 static moduledata_t cxgbei_mod = {
1019 	"cxgbei",
1020 	cxgbei_modevent,
1021 	NULL,
1022 };
1023 
1024 MODULE_VERSION(cxgbei, 1);
1025 DECLARE_MODULE(cxgbei, cxgbei_mod, SI_SUB_EXEC, SI_ORDER_ANY);
1026 MODULE_DEPEND(cxgbei, t4_tom, 1, 1, 1);
1027 MODULE_DEPEND(cxgbei, cxgbe, 1, 1, 1);
1028 MODULE_DEPEND(cxgbei, icl, 1, 1, 1);
1029